C-BRIC Researchers Present at ECCV Conference

C-BRIC researchers presented their work at the 2022 European Conference on Computer Vision (ECCV). The conference was held October 23-27, 2022, in Tel Aviv, Israel.

 

Purdue researchers Sayeed Chowdhury, Nitin Rathi, and Kaushik Roy; University of Southern California researchers Yunhao Ge, Amanda S. Rios, and Laurent Itti; and Yale researchers Youngeun Kim, Yuhang Li, Hyoungseob Par, Yeshwanth Venkatesha, Tamar Geller, Ruokai Yin, and Priya Panda represented C-BRIC well at ECCV.

 

The Purdue research team (Chowdhury, Rathi, and Roy) presented their work on "One Timestep is All You Need: Training Spiking Neural Networks with Ultra Low Latency." In this work, they proposed the Iterative Initialization and Retraining method for SNNs (IIR-SNN). The method starts with an SNN trained with T timesteps (T>1). Then at each stage of latency reduction, the network trained at the previous stage with a higher timestep is utilized as initialization for subsequent training with a lower timestep. This is a compression method that achieves top accuracy, reduces latency, and provides higher energy efficiency. Sayeed Chowdhury is currently a PhD student at Purdue University under the direction of Kaushik Roy. Nitin Rathi is a PhD alum from Roy's group, currently with Facebook. Roy is the C-BRIC Director and a faculty member at Purdue. 

 

The USC research group, including Ge, Itti, and their USC collaborators Zhi Xu, Yao Xiao, and Xingrui Wang, presented "Contributions of Shape, Texture, and Color in Visual Recognition." This paper investigated the contributions of three important features of the human visual system (HVS) --- shape, texture, and color --- to object classification. They built a humanoid vision engine (HVE) that explicitly and separately computes shape, texture, and color features from images and showed that HVE can summarize and rank-order the contributions. HVE can be used to simulate the open-world zero-shot learning ability of humans with no attribute labeling and show that HVE can also simulate human imagination ability with the combination of different features. Yunhao Ge is a PhD student at USC under the direction of Laurent Itti. Itti is a faculty member at USC.

 

USC researchers Rios and Itti and Intel collaborators Ergin Genc, Ibrahima Ndiour, Nilesh Ahuja, and Omesh Tickoo presented "incDFM: Incremental Deep Feature Modeling for Continual Novelty Detection." The paper proposed incDFM (incremental Deep Feature Modeling), a self-supervised continual novelty detector. The method builds a statistical model over the space of intermediate features produced by a deep network and utilizes feature reconstruction errors as uncertainty scores to guide the detection of novel samples. incDFM estimates the statistical model incrementally and selects only the most confident novel samples each time, which will then guide subsequent recruitment incrementally to achieve state of the art continual novelty detection performance. Amanda Rios is an alumna of USC who studied under the direction of Laurent Itti. 

 

USC researchers Ge and Itti and their collaborators Jiashu Xu of USC, Harkirat Behl, Suriya Gunasekar, Neel Joshi, Yale Song, Xin Wang, and Vibhav Vineet of Microsoft presented "Neural-Sim: Learning to Generate Training Data with NeRF." This work presented the first fully differentiable synthetic data generation pipeline that uses Neural Radiance Fields (NeRFs) in a closed-loop with a target application's loss function to generate data, on demand, with no human labor, to maximize accuracy for a target task. The group illustrated the method's effectiveness with synthetic and real-world object detection experiments and evaluated it on a new "YCB-in-the-Wild" dataset that provides a test scenario for object detection with varied poses in real-world environments.

 

The Yale team of Youngeun Kim, Yuhang Li, Hyoungseob Park, Yeshwanth Venkatesha, and Priya Panda presented "Neural Architecture Search for Spiking Neural Networks." This paper introduced a novel Neural Architecture Search (NAS) approach for finding better SNN architectures. SNASNet, found by the team's search algorithm, achieves higher performance with backward connections, demonstrating the importance of designing SNN architecture for suitably using temporal information. It also achieves state-of-the-art performance with significantly lower timesteps. Kim, Li, Park, and Venkatesha are all current PhD students under the direction of Priya Panda. Panda is a faculty member at Yale.

 

Yale researchers Yuhang Li, Youngeun Kim, and Priya Panda presented "Neuromorphic Data Augmentation for Training Spiking Neural Networks" as well. In this paper, the group proposed Neuromorphic Data Augmentation (NDA), a family of geometric augmentations specifically designed for event-based datasets with the goal of significantly stabilizing the SNN training and reducing the generalization gap between training and test performance. The proposed method is simple and compatible with existing SNN training pipelines. The proposed augmentation demonstrated the feasibility of unsupervised contrastive learning for SNNs.

 

A third group of C-BRIC researcher at Yale also presented at ECCV. Youngeun Kim, Yuhang Li, Hyoungseob Park, Yeshwanth Venkatesha, Priya Panda, and Ruokai Yin presented "Lottery Ticket Hypothesis for Spiking Neural Networks." This work investigated Lottery Ticket Hypothesis (LTH) which states that dense networks contain smaller subnetworks (i.e., winning tickets) that achieve comparable performance to the dense networks. Our studies on LTH reveal that the winning tickets consistently exist in deep SNNs across various datasets and architectures, providing up to 97% sparsity without huge performance degradation. Yin is a PhD student under the direction of Priya Panda.